Should we return Symmetric matrices?

[nickrobinson251]

No description provided.

[tlienart]

That seems reasonable, IIRC in a lot of places the estimators are based upon calls to the default cov (and linear combinations thereof) and so properties should (?) be preserved, though we also compute the uncentered sample covariance which should be wrapped in a Symmetric wrapper and all this should be tested.

@mateuszbaran thoughts? also, I'm potentially happy to PR for this (unless @nickrobinson251 you wanted to take the lead?)

[nickrobinson251]

Cool - go for it! (if @mateuszbaran is onboard ofc) -- i'd be happy to review :)

[mateuszbaran]

That's a good idea, although I'd be careful implementing this. Standard cov works for complex random variables but then the matrix is Hermitian. I haven't tested which of our estimators behave correctly for complex numbers (there is a separate issue about it, #29), SimpleCovariance definitely works though and some linear shrinkage estimators might work with minor modifications. The point is, if we do this, we should wrap all matrices in either Symmetric or Hermitian and SimpleCovariance needs to cover both cases.

I'd be happy to review a PR 🙂 .

[mateuszbaran]

Fixed by #64.